DeepSORT Pedestrian Tracking Algorithm based on Azimuth Estimation

In order to solve the problem that pedestrian tracking accuracy is not high and identity exchange is frequent in complex road environments, this paper proposes an improved DeepSORT target tracking algorithm combining azimuth estimation and feature moving average (FMA) feature association strategy. The algorithm uses Kalman filter algorithm to predict and update the detection results in real-time, so as to realize the classification and tracking of the same category in the video. To solve the problem of feature extraction in the tracking process, this paper adopts the FMA to extract the features of the current frame detection frame and calculate the average value of the previous frame features, so as to improve the correlation degree of features between video frames. At the same time, in order to better locate the position information of the detection frame and the prediction frame in the process of cascade matching, this paper puts forward the azimuth estimation and applies it to the cost matrix to make the detection frame match the optimal prediction. The experimental results show that MOTA, MOTP and other indexes of the proposed algorithm are greatly improved, and the frequency of identity exchange is also reduced to a certain extent.